(IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 14, No. 6, 2023 An Enhanced Variational AutoEncoder Approach for the Purpose of Deblurring Bangla License Plate Images Md. Siddiqure Rahman Tusher, Nakiba Nuren Rahman, Shabnaz Chowdhury, Anika Tabassum, Md. Akhtaruzzaman Adnan, Rashik Rahman * , Shah Murtaza Rashid Al Masud * Department of Computer Science and Engineering University of Asia Pacific, Dhaka, Bangladesh Abstract—Automated License Plate Detection and Recognition (ALPDR) is a well-studied area of computer vision and a crucial activity in a variety of applications, including surveillance, law enforcement, and traffic management. Such a system plays a crucial role in the investigation of vehicle-related offensive activities. When an input image or video frame travels through an ALPDR system for license plate detection, the detected license plate is frequently blurry due to the fast motion of the vehicle or low-resolution input. Images of license plates that are blurred or distorted can reduce the accuracy of ALPDR systems. In this paper, a novel Variational AutoEncoder(VAE) architecture is proposed for deblurring license plates. In addition, a dataset of obscured license plate images and corresponding ground truth images is proposed and used to train the novel VAE model. This dataset comprises 3788 image pairs, in which the train, test, and validation set contains 2841, 568, and 379 pairs of images respectively. Upon completion of the training process, the model undergoes an evaluation procedure utilizing the validation set, where it achieved an SSIM value of 0.934 and a PSNR value of 32.41. In order to assess the efficacy of our proposed VAE model, a comparison with contemporary deblurring techniques is pre- sented in the results section. In terms of both quantitative metrics and the visual quality of the deblurred images, the experimental results indicate that our proposed method outperforms the other state-of-the-art deblurring methods. Therefore, it enhances the precision and dependability of an ALPDR system. Keywords—Image deblur; bangla license plate deblur; Varia- tional AutoEncoder (VAE); computer vision I. I NTRODUCTION With the increase in the number of vehicles on the road, violations of traffic laws such as racing through red lights, leaving the scene of an accident, and kidnapping escalated. As a result, ALPDR systems have been extensively developed and applied to a variety of intelligent traffic systems [1]– [4]. Unfortunately, despite the fact that drive recorders and surveillance cameras perform significantly better than in the past, license plates of vehicles are frequently blurred due to fast-moving vehicles, camera shakes during the exposure period, and other factors. Multiple variables contribute to the blurring and distortion of license plate images. The initial variable is the local environment. For instance, the impacts of intense illumination, precipitation, and weather may increase the likelihood of blurring. The second variable consists of the vehicle’s motions. For instance, when vehicles run red lights, they are frequently traveling at a very high rate of speed; consequently, the pictures taken tend to be blurry. The surveillance system serves as the final variable. Due to the fact that surveillance cameras are frequently positioned at higher elevations, far from the car, the captured image has a reduced resolution, leading to poor image quality. Blurred license plate images can significantly reduce the accuracy of ALPDR systems. Therefore, deblurring of the license plate images is a crucial step towards achieving reliable ALPDR systems. Starting from deblurring images using methods such as dihedral group [5], image deblurring techniques have signif- icantly advanced [6] in recent years. On the basis, of high- frequency residual image learning, the authors of [7] proposed a two-phase deblurring algorithm for restoring blurred images of dynamic scenes. The method proposed in [8] defines a new regularization term that incorporates both intensity and gradient assumptions and provides an efficient and convergent solution for deblurring license plate images. Convolutional Neural Networks (CNNs) have been extensively utilized [9]– [11] alongside Generative Adversarial Networks (GANs) [8], [12] to generate sharp license plate images. However, these approaches demand a large amount of training data, as with a small dataset the model may not converge and can lead to a less generalized model. To our knowledge, there are a handful of works related to Bangla license plate deblurring. The purpose of this study is to develop and establish a state-of-the-art Bangla license plate deblurring system that can work in real-time. In order to accomplish this we propose a novel VAE architecture with a custom loss function for Bengali license plate image deblurring. The proposed VAE network is a generative model capable of learning the underlying distribution of training data and producing new samples based on the learned distribution. Thus, it can perform well even if it is trained on a small dataset. However, there’s a lack of publicly available Bangla license plate dataset for deblurring. Thus, we created a new balanced and generalized dataset consisting of 3788 pairs of images that were used to train, test and validate the model. 75% of the data belong to the train set, whereas 15% and 10% data belong to the test and validation sets, respectively. The proposed VAE model achieved an SSIM score of 0.934 and a PSNR score of 32.41. To assess the efficacy, we recreated state-of-the-art deblurring models [13], [14] and trained them on our dataset. The evaluation demonstrates that our method outperforms state- of-the-art deblurring techniques in terms of both quantitative metrics and image quality. In addition, we demonstrate the www.ijacsa.thesai.org 1252 | Page